nlst_sybil / cropping_analysis /launch_parallel_slice_drop.py
chn123's picture
Sanitize absolute server paths to relative paths
f182a8d verified
Raw History Blame Contribute Delete
4.2 kB
#!/usr/bin/env python
import os
import sys
import time
import json
import subprocess
from pathlib import Path
PYTHON_EXEC = "./anaconda3/envs/sybil_python311_env/bin/python"
WORKER_SCRIPT = "./examples/slice_drop_worker.py"
CHUNK_DIR = "./examples/slice_drop_chunks"
FINAL_JSON = "./examples/slice_drop_all_2965_results.json"
NUM_WORKERS = 3
GPU_MAP = ["cuda:0", "cuda:1", "cuda:2"] # Mapped via CUDA_VISIBLE_DEVICES=0,1,3
def start_workers():
os.makedirs(CHUNK_DIR, exist_ok=True)
processes = []
env = os.environ.copy()
env["CUDA_VISIBLE_DEVICES"] = "0,1,3"
for wid in range(NUM_WORKERS):
log_path = os.path.join(CHUNK_DIR, f"worker_{wid}.log")
log_file = open(log_path, "a")
cmd = [
PYTHON_EXEC, WORKER_SCRIPT,
"--worker_id", str(wid),
"--num_workers", str(NUM_WORKERS),
"--device", GPU_MAP[wid],
"--chunk_dir", CHUNK_DIR
]
print(f"Launching Worker {wid} on {GPU_MAP[wid]} (Logs: {log_path})...")
p = subprocess.Popen(cmd, env=env, stdout=log_file, stderr=subprocess.STDOUT)
processes.append((wid, p, log_file))
return processes
def get_progress():
counts = []
for wid in range(NUM_WORKERS):
out_file = os.path.join(CHUNK_DIR, f"worker_{wid}_results.json")
if os.path.exists(out_file):
try:
with open(out_file, "r") as f:
data = json.load(f)
counts.append(len(data))
except Exception:
counts.append(0)
else:
counts.append(0)
return counts
def merge_results():
all_results = {}
for wid in range(NUM_WORKERS):
out_file = os.path.join(CHUNK_DIR, f"worker_{wid}_results.json")
if os.path.exists(out_file):
try:
with open(out_file, "r") as f:
data = json.load(f)
for item in data:
all_results[item["series_id"]] = item
except Exception as e:
print(f"Merge error on worker {wid}: {e}")
merged_list = list(all_results.values())
with open(FINAL_JSON, "w") as f:
json.dump(merged_list, f, indent=2)
return len(merged_list)
def monitor_workers(processes, total_target=2965):
start_time = time.time()
last_print = 0
print("\nAll workers started! Monitoring execution...")
try:
while True:
time.sleep(15)
counts = get_progress()
total_done = sum(counts)
# Check if any process died unexpectedly
all_alive = True
for wid, p, _ in processes:
poll = p.poll()
if poll is not None and poll != 0:
print(f"WARNING: Worker {wid} exited with code {poll}!")
all_alive = False
now = time.time()
if now - last_print >= 60 or total_done >= total_target or not all_alive:
last_print = now
elapsed = now - start_time
speed = total_done / elapsed if elapsed > 0 else 0
eta_sec = (total_target - total_done) / speed if speed > 0 else 0
eta_hours = eta_sec / 3600
pct = (total_done / total_target) * 100
print(f"[{time.strftime('%X')}] Progress: {total_done}/{total_target} ({pct:.1f}%) | "
f"Workers: {counts} | Speed: {speed*60:.1f} series/min | ETA: {eta_hours:.1f}h")
sys.stdout.flush()
# Update merged json periodically
merged_count = merge_results()
# If all processes finished
if all(p.poll() is not None for _, p, _ in processes):
print("\nAll worker processes completed.")
break
finally:
merge_results()
for _, _, log_file in processes:
log_file.close()
if __name__ == "__main__":
procs = start_workers()
monitor_workers(procs)